An experiment was performed to test the effectiveness of a multivariate method of analysis for distinguishing earthquakes from explosions. The data base for the experiment consisted of digital recordings made at twenty- seven stations, including four large arrays, of one hundred thirty-three Eurasian events. Spectral magnitudes were measured in three frequency bands of the six phases P, Lg, long-period P, long-period S, LR, and LQ. Complexities of the P-waves were measured in three time windows, and corner frequencies and low- frequency spectral levels were computed for the P-wave spectra. When signals could not be detected, spectral magnitudes were measured of noise samples, and these were taken to be detection thresholds. Signal magnitudes and detection thresholds from all the stations were used to find maximum-likelihood estimates of the spectral magnitudes for each event. On account of gaps in the data, many different combinations of different numbers of variables were used in the classification of the one hundred thirty-three events. The variables which were applicable to the most events were short-period P-wave spectral ratios. The most effective discriminants were spectral ratios of Love waves to P waves.